A CSV column can look categorical while carrying a mixture of observations, conditions, and reported claims. The watermark column in StagingBench is a useful example. Treating it as a ready made Boolean would discard information before the analysis even starts.
StagingBench is an independent test of AI virtual staging tools on 10 real room photos, with one run per tool and room cell. The timing and watermark table used here contains 112 rows from the experiment's 120 cells. Its watermark text is therefore a record for the rows present, not complete coverage of every cell.
Start with the actual vocabulary
There are 19 distinct strings in the column. Two exact examples show why inspecting them matters:
NoNo visible watermark (complimentary Pro)
The first string is short. The second adds a condition about the output being described. Other entries are longer sentences that combine a service's statement about its free plan with an observation about a saved preview. Those entries do not begin with Yes or No. Assigning them to a binary category would require an interpretive rule beyond the literal opening of the text.
For a transparent first count, group by the leading word, case insensitively: Yes, No, or anything else. Applied to the supplied values, that gives 56 Yes rows, 46 No rows, and 10 other rows. These groups account for all 112 rows in the table.
Make the counting rule visible
The following snippet reads timing-watermark.csv in the current directory. It counts text prefixes, preserves the original CSV, and prints every group, including the residual category.
import csv counts = {"Yes": 0, "No": 0, "other": 0} with open("timing-watermark.csv", newline="", encoding="utf-8") as handle: reader = csv.DictReader(handle) for row in reader: text = row["watermark"].casefold() if text.startswith("yes"): counts["Yes"] += 1 elif text.startswith("no"): counts["No"] += 1 else: counts["other"] += 1 for label, count in counts.items(): print(f"{label}: {count}") I ran this code against the table. Its actual output was:
Yes: 56 No: 46 other: 10 The code implements a prefix test. For the strings actually present, its result agrees with inspecting the leading words. That agreement depends on checking the vocabulary first. A future entry beginning with a longer word that merely shares a prefix could be counted differently from what a reader expects.
Why inspection comes before aggregation
Free text often grows through practical note taking. A short answer acquires a qualification, a preview differs from a saved download, or an observer records what a service says alongside what they saw. Counting requires choosing which of those distinctions to preserve.
Case insensitivity handles capitalization variation without pretending that sentences have become standardized categories. It does not resolve conflicting claims within a sentence. Nor does the other category mean that a watermark was present, absent, or impossible to assess. It means only that the entry did not start with the selected prefixes.
A useful workflow is to inspect distinct values, choose an explicit rule, run the count, and examine what remains outside the main categories. Keeping that residual visible gives a later reviewer somewhere to look when deciding whether a richer classification is justified.
Keep the conclusion attached to the evidence
These totals describe textual records. They are not a fresh visual inspection of the staged photographs, and they do not establish what a future account or download would contain. Conditions embedded in a note remain relevant even when its opening word places it in a simple group.
The StagingBench methodology links to the benchmark's published explanation. For this small test, the defensible result is a reproducible count with a visible rule. With 10 rooms and one run per cell, extending that count into a general promise about a service would require evidence beyond this file.

